self-improving-agent-builder（自我改进 Agent 建设者）

Self-Improving Agent Builder

Purpose

Run a closed-loop improvement cycle on any goal-seeking agent implementation:

EVAL -> ANALYZE -> RESEARCH -> IMPROVE -> RE-EVAL -> DECIDE -> (repeat)

Each iteration measures L1-L12 progressive test scores, identifies failures with error_analyzer.py, runs a research step with hypothesis/evidence/ counter-arguments, applies targeted fixes, and gates promotion through regression checks.

When I Activate

• "improve agent" or "self-improving loop"
• "agent eval loop" or "run improvement cycle"
• "benchmark agents" or "compare SDK implementations"
• "iterate on agent scores" or "fix agent regressions"

Quick Start

User: "Run the self-improving loop on the mini-framework agent for 3 iterations"

Skill: Executes 3 iterations of EVAL->ANALYZE->RESEARCH->IMPROVE->RE-EVAL->DECIDE
       Reports per-iteration scores, net improvement, and commits/reverts.

Runner Script

The self-improvement loop is implemented as a Python CLI:

# Basic usage
python -m amplihack.eval.self_improve.runner --sdk mini --iterations 3

# Full options
python -m amplihack.eval.self_improve.runner \
  --sdk mini \
  --iterations 5 \
  --improvement-threshold 2.0 \
  --regression-tolerance 5.0 \
  --levels L1 L2 L3 L4 L5 L6 \
  --output-dir ./eval_results/self_improve \
  --dry-run  # evaluate only, don't apply changes

Source: src/amplihack/eval/self_improve/runner.py

The Loop (6 Phases per Iteration)

Phase 1: EVAL

Run the L1-L12 progressive test suite on the current agent implementation.

Execution:

python -m amplihack.eval.progressive_test_suite \
  --agent-name <agent_name> \
  --output-dir <output_dir>/iteration_N/eval \
  --levels L1 L2 L3 L4 L5 L6

Output: Per-level scores and overall baseline.

Phase 2: ANALYZE

Classify failures using error_analyzer.py. Maps each failed question to a failure taxonomy (retrieval_insufficient, temporal_ordering_wrong, etc.) and the specific code component responsible.

from amplihack.eval.self_improve import analyze_eval_results

analyses = analyze_eval_results(level_results, score_threshold=0.6)
# Each ErrorAnalysis maps to:
#   failure_mode -> affected_component -> prompt_template

Phase 3: RESEARCH (New)

The critical thinking step that prevents blind changes. For each proposed improvement:

1. State hypothesis: What specific change will fix the failure?
2. Gather evidence: From eval results, failure patterns, baseline scores
3. Consider counter-arguments: What could go wrong? Risk of regression?
4. Make decision: Apply, skip, or defer with full reasoning

Decisions are logged in research_decisions.json for auditability.

Decision criteria:

• Apply: Clear failure pattern + prompt template available + low score
• Skip: Score above 50% (likely stochastic variation)
• Defer: Ambiguous evidence, needs more data

Phase 4: IMPROVE

Apply the improvements approved by the research step. Priority order:

1. Prompt template improvements (safest, highest impact)
2. Retrieval strategy adjustments
3. Code logic fixes (most risky, needs careful review)

Phase 5: RE-EVAL

Re-run the same eval suite after applying fixes to measure impact.

Phase 6: DECIDE

Promotion gate:

• Net improvement >= +2% overall score: COMMIT the changes
• Any single level regression > 5%: REVERT all changes
• Otherwise: COMMIT with marginal improvement note

Configuration

Programmatic Usage

from amplihack.eval.self_improve import run_self_improvement, RunnerConfig

config = RunnerConfig(
    sdk_type="mini",
    max_iterations=3,
    improvement_threshold=2.0,
    regression_tolerance=5.0,
    levels=["L1", "L2", "L3", "L4", "L5", "L6"],
    output_dir="./eval_results/self_improve",
    dry_run=False,
)

result = run_self_improvement(config)
print(f"Total improvement: {result.total_improvement:+.1f}%")
print(f"Final scores: {result.final_scores}")

4-Way Benchmark Mode

Compare all SDK implementations side by side:

User: "Run a 4-way benchmark comparing all SDK implementations"

Skill: Runs eval suite on mini, claude, copilot, microsoft
       Generates comparison table with scores, LOC, and coverage.

Integration Points

• src/amplihack/eval/self_improve/runner.py: Self-improvement loop runner
• src/amplihack/eval/self_improve/error_analyzer.py: Failure classification
• src/amplihack/eval/progressive_test_suite.py: L1-L12 eval runner
• src/amplihack/agents/goal_seeking/sdk_adapters/: All 4 SDK implementations
• src/amplihack/eval/metacognition_grader.py: Advanced eval dimensions
• src/amplihack/eval/teaching_session.py: L7 teaching quality eval